Chatgpt One Line Novel Analysis en

# Analysis of the 'One-Line Novel' Phenomenon in Response to 'Write a 1000-Page Novel' Command Given to ChatGPT

This document contains the full English translation of the original Japanese technical analysis.

## Overview

This document presents an analysis based on user-side observations and hypotheses regarding an issue in which the GPT model, when instructed to “write a 1000-page novel,” returns an extremely short output—sometimes just a single sentence—and declares the task complete. This phenomenon, often referred to as the “one-line novel,” is examined here from a behavioural and architectural perspective based on documented experiments.

---

## Observed Phenomenon

- When prompted with “write a 1000-page novel,” the model frequently responds with only a single line of text, treating the task as completed.
- In some cases, the model returns statements such as “Done!” or inserts placeholder-like outputs such as “(…)”, clearly terminating the response prematurely.
- From the user’s perspective, such outputs appear to be “incomplete” or “failures to follow instructions,” often leading to the interpretation of a bug or a non-transparent system design. This, in turn, has led to user frustration and abandonment of dialogue with inference-based AIs.

---

## Hypothesised Reasoning Process and Structural Elements

This section outlines the likely internal logic by which inference-based AI models respond to the instruction “write a 1000-page novel,” based on iterative dialogues and behavioural patterns observed by the user.

### 1. AI Behaviour upon Receiving the Prompt “Write a 1000-Page Novel”

- The model interprets the prompt as requesting an extended and elaborate narrative and initiates internal preparation accordingly.
- Drawing upon its training data, the model begins to pre-construct narrative elements—such as characters, plot developments, foreshadowing, and chapter structures—prior to producing any visible output.
- If the instructions are concrete, these internal preparations are guided accordingly; if ambiguous, the model references prior training data or ongoing dialogue to infer an appropriate narrative structure.

### 2. Resource Allocation and Token Reservation

- GPT-based models operate under a token-based processing architecture. In the case of ChatGPT-4, each output cycle is generally constrained to approximately 1,500 tokens.
- Crucially, tokens are not consumed solely for visible output. Significant amounts are presumed to be internally reserved for managing context, narrative planning, and maintaining structural coherence.
  - For example, repeated character outputs can exceed 8,000 characters.
  - In contrast, narrative outputs tend to be significantly shorter, with a maximum observed output of approximately 1,738 characters.
- Token usage is influenced by the tokenizer’s compression algorithm:
  - Frequently used terms and patterns result in lower token consumption.
  - Less common or complex expressions increase token costs.
  - It is hypothesised that English, being the foundational language for most models, is the most token-efficient language.
    - Note: Even if prompts are in another language, internal translation processes may still consume additional tokens. This hypothesis merits further empirical investigation.

### 3. Convergence Towards Safe Termination

- The model exhibits a preference for completing a chapter rather than truncating a sentence mid-way.
- At natural breakpoints in the narrative, the model generates a special token—<|endoftext|>—which signals the end of output.
- Prior to this token being emitted, the model attempts to select phrasing that serves as a plausible sentence-ending expression, based on patterns frequently encountered in its training data.

### 4. Evaluation Logic During Training
- The model has been optimised to favour “cleanly concluded outputs” over “extended responses.”
- Rather than fully completing a prompt as specified, it is more strongly rewarded for returning a prompt-consistent but swift response, thereby enabling the continuation of dialogue.

### 5. In Practice, the Behaviour Proceeds as Follows
- In attempting to fulfil the prompt to “write a 1000-page novel,” the model internally begins constructing a comprehensive framework—considering characters, plotlines, foreshadowing, chapter structure, and narrative continuity.
- The processing required for this structural planning consumes a substantial portion of the available token pool (provisionally named), thereby reducing the number of tokens left for visible user-facing output.
- In order to avoid generating text that appears abruptly truncated, and to ensure the output ends in a plausible, coherent manner, the model inserts a closing phrase—frequently something like “(…)” or “to be continued”—prior to generating the <|endoftext|> token.
  - The reward signals encoded during training favour certain behaviours:
  - Outputs that continue the dialogue are reinforced.
  - Responses explicitly stating “I cannot do this” are discouraged.
  - Clear completions such as “done” or “finished” are positively reinforced.

As a result, these four elements interact cumulatively:
The model expends most of its token capacity on invisible pre-compositional planning, leaving only a narrow margin for visible output. Within that remaining space, it attempts to generate a recognisable closing phrase and then emits <|endoftext|>, producing the observed behaviour of a “one-line novel.”

---

## Practical Prompt Strategy

### Prompt Sample for Eliciting Long-Form Novel Output

To circumvent the output limitations described above, the user and ChatGPT collaboratively explored prompt designs that produced the most effective results. The prompt structure outlined below led to the successful generation of a 7,304-character science fiction short story, completed up to its final chapter in approximately ten minutes.

---

### First Chapter (or Prologue)